@LotfiSanae on Backscroll

2 tweets from @LotfiSanae that Backscroll ranked for substance.

Sanae Lotfi @ICML @LotfiSanae 2026-07-06 21:28Z

Very nice to see our recently proposed overthinking penalty tested and working for others, leading to ~30% cot length reduction here! As we mention in the paper, the penalty on overthinking markers improves efficiency for both quantized and full precision models without hurting

Sanae Lotfi @LotfiSanae 2026-06-25 17:39Z

When a quantized reasoning model fails, we usually assume it just lost the capability to solve the problem. But we surprisingly show that in up to 52% of failures, the quantized model actually reaches the correct answer but then does not commit to it. 1/8

How it ranks: Backscroll reads my Twitter/X timeline and ranks every tweet for substance with an LLM taste rubric — not raw engagement. Curated by Surya Dantuluri.